Towards the Design of Robust Trust and Reputation Systems

1 Research Problem

Our research is within the area of artificial intelligence and multiagent system. More specifically, we are interested in addressing robustness problems in trust and reputation systems so that buying agents are able to accurately model the reputation of selling agents even with the existence of various unfair rating attacks from other dishonest buyers (called advisors) [Zhang and Cohen, 2008].

In multiagent-based e-marketplaces, trust and reputation systems are designed for buyers to model seller reputation based on ratings shared by advisors. However, unfair rating attacks from dishonest advisors render trust and reputation systems ineffective to mislead buyers to transact with dishonest sellers [Jøsang, 2012]. Typical unfair rating attacks include Constant where dishonest advisors constantly provide unfairly positive/negative ratings to sellers; Camouflage where dishonest advisors camouflage themselves as honest advisors by providing fair ratings to build up their trustworthiness first and then gives unfair ratings; Whitewashing where a dishonest advisor is able to whitewash its low trustworthiness by starting a new account with the initial trustworthiness value; Sybil where a dishonest buyer creates several accounts to constantly provide unfair ratings to sellers; Sybil Camouflage (the combination of Sybil and Camouflage) where a number of dishonest advisors perform Camouflage attacks together; and Sybil Whitewashing where a number of dishonest advisors performWhitewashing attacks together. The challenge is thus how to design robust trust models to against various strategic attacks.

Various trust models [Zhang and Cohen, 2008; Jøsang, 2012] have been proposed to cope with unfair ratings. However, these models are not completely robust against various attacks. In particular, when dishonest advisors occupy a large proportion in e-marketplaces (i.e., Sybil), BRS becomes inefficient and iCLUB is unstable because they both employ the "majority-rule". When dishonest advisors adopt strategic attacks, TRAVOS does not work well because it assumes an advisors' rating behavior is consistent. ReferralChain assigns trust value 1 to every new buyer (advisor) which provides a chance for dishonest advisors to abuse the initial trust (i.e., Whitewashing). Personalized is vulnerable when buyers have insufficient experience with advisors and the majority of advisors are dishonest (i.e., combination of Whitewashing and Sybil). Thus, we need more robust trust models.

From results in [Zhang et al., 2012], the robustness of single trust models can be enhanced by combining different categorial models, and Discount-then-Filter is most robust than Filter-then-Discount against typical unfair rating attacks.

2.2 Construction of Robust Trust Network

To date, most trust models are designed based on statistical techniques. In [Jiang et al., 2013], we propose a novel multiagent evolutionary trust model (MET) where each buyer constructs its trust network (information about which advisors should be include in the network and their trustworthiness) by the evolutionary model.

Assume that in an e-marketplace, the set of buyers is denoted as B = {Bi|i = 1, · · · , l} and the set of sellers is denoted as S = {Sj|j = 1, · · · ,m}. We also denote the trustworthiness of an advisor Ak ∈ B from the view of a buyer Bi as TBi (Ak) ∈ [0, 1]. In the buyer Bi's trust network TNBi , the trustworthiness values of advisors connected with Bi is then denoted as TBi (A) = {TBi (Ak)|Ak ∈ TNBi }.

In MET, a fitness function is designed for buyers to measure the quality of their trust networks. Formally, the fitness value of buyer Bi's trust network TBi (A) is calculated as:

where RBi (Sj) and ˜RBi (Sj) are reputation of a seller Sj based on Bi's personal experience and advisors' ratings from TBi (A), respectively. In addition, m′ ≤ m, indicating that sellers with which either buyer Bi or its advisors have no experience will not be considered in the fitness evaluation.

A smaller fitness value indicates that the buyer's trust network is in higher quality, because the combination of advisors' ratings is more similar to the buyer's own opinions regarding common sellers. In other words, the fitness function measures the suitability of selected advisors and the accuracy of the trust values assigned to these advisors simultaneously.

In each generation, buyer Bi choose advisor Ar to take interaction only when the following condition is satisfied:

After choosing three 1 advisors by trust network comparison, the buyer will generate a candidate trust network using evolutionary operators [Jiang et al., 2012]. Two widely used evolutionary operators (DE crossover and polynomial mutation) are adopted to produce candidate trust works in MET. By comparing the candidate trust network with the buyer's own trust network, the one with higher fitness value measured by Eq. 2 will survive to the next generation.

From Table 1, experimental results show that iCLUB and the Personalized approach have large perturbation under Sybil attacks. BRS, TRAVOS and ReferralChain are vulnerable to Sybil, Camouflage andWhitewashing, respectively. It demonstrates that MET is more robust than these other trust models against typical unfair rating attacks.

1 If insufficient advisors satisfy Eq. 3, some advisors will be randomly selected from the buyer's trust network.

Table 1: Robustness of Trust Models vs. Attacks

∗ Sybil Cam: Sybil Camouflage; Sybil WW: Sybil Whitewashing

3 Future Research

In future, we will investigate the impact of various parameters and components on our trust models, i.e., the threshold in Discount-then-Filter, different types of evolutionary operators inMET, etc. In our METmodel, the trust network information for buyers is encoded as the fixed length of trust values for advisors. We plan to test different encoding methods, such as varying length of trust networks for different buyers and graph encoding to represent buyers' trust networks.

We also plan to build a comprehensive testbed to evaluate the robustness of trust models based on simulated and real e-commerce environments. In the current work, we only investigated the typical unfair rating attacks from advisors. In future, we will incorporate more intelligent attacks from advisors. For instance, dishonest advisors firstly adopt learning techniques to analyze a buyer's rating pattern, and then inject a small number of unfair ratings into e-marketplaces to mislead the buyer to transact with dishonest sellers. We will also combine sellers' cheating behaviors with advisors' unfair ratings, and evaluate their impact on different trust models.